DF-Net: The Digital Forensics Network for Image Forgery Detection

TL;DR

DF-Net tackles image forgery detection with a focus on robustness to lossy OSN operations. It employs two U-Net-based sub-networks with scSE blocks to perform pixel-wise forgery localization, fused by a per-pixel maximum to produce a final manipulation likelihood map. Trained on the DF2023 dataset, DF-Net achieves state-of-the-art performance across four benchmarks and demonstrates strong resilience to OSN-induced distortions, while offering faster inference than tiling-based methods. The work provides an open-source detector suitable for real-world forensic applications across multiple manipulation types.

Abstract

The orchestrated manipulation of public opinion, particularly through manipulated images, often spread via online social networks (OSN), has become a serious threat to society. In this paper we introduce the Digital Forensics Net (DF-Net), a deep neural network for pixel-wise image forgery detection. The released model outperforms several state-of-the-art methods on four established benchmark datasets. Most notably, DF-Net's detection is robust against lossy image operations (e.g resizing, compression) as they are automatically performed by social networks.

Figures (5)

• Figure 1: Forgery detection results of our network. Example images are taken from the CASIACASIA and the NISTNIST datasets.
• Figure 2: Network architecture of submodel M1: A U-Net architecture with 4 skip connections and spatial channel Squeeze & Excitation (scSE) extension. A more detailed description can be found in section \ref{['sec:architecture']}.
• Figure 3: Network architecture of DF-Net: Example of Model combination for image Sp_S_NNN_C_txt0019_txt0019_0019.jpg from the CASIA_V1 dataset.
• Figure 4: Metrics (AUC, F1, IoU) averaged over 4 benchmark datasets in accumulated presentation. Each column represents the combination of a method and the OSN used for dataset modification
• Figure 5: Examples of qualitative comparison of MantraNet Wu2019, Wu22 Wu2022 and our proposed DF-Net. Each line shows one example image for each of the four benchmark datasets DSODSO, ColumbiaColumbia, NISTNIST, CASIACASIA. The five columns show: the forged image (input), manipulated area (ground truth), results (output) from MantraNet, Wu22 and DF-Net. We show example results of the M1 sub-model for the DSO and the NIST dataset.